Face Verification Using ConvNet Hidden Identity Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing face verification methods rely on shallow models and low-level features, which are not sufficiently discriminative for fine-grained classes of face identities, and often lose useful information during feature extraction.
Innovation Solution
The use of different trained ConvNets to extract Hidden Identity Features (HIFs) from various face regions, with the last hidden layer neuron activations forming feature vectors that are concatenated and compared to determine identity matching, employing a training process that focuses on identity classification and utilizes dropout for robust feature learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If shallow face verification models with low-level features are used, then the system complexity is low, but the discrimination capability for fine-grained face identities is insufficient
Solution Approach 1:
The face image is divided into multiple local regions, and separate ConvNets are trained for each region to extract specialized features. This segmentation allows the system to capture fine-grained identity information from different facial parts while maintaining manageable model complexity through modular architecture.
Solution Approach 2:
The patent transitions from shallow feature extraction to deep hierarchical feature extraction using multiple convolutional layers. This dimensional transformation in feature space enables the model to learn complex identity representations that go beyond low-level features, achieving better discrimination capability.
2Reliability
If face similarity features are extracted from a pair of faces using multiple deep ConvNets, then the features are highly discriminative, but useful information may be lost before final verification
Solution Approach 1:
The patent performs preliminary feature extraction from individual face images before comparison. By extracting features from each face separately and then combining them, the system preserves more information than direct pair-wise extraction, as each face's unique characteristics are captured independently before verification.
Solution Approach 2:
Features extracted from multiple local face regions are concatenated to form a comprehensive feature vector. This merging of regional features preserves information from different facial parts while maintaining high discriminability through the combined representation.
3Device complexity
If the last hidden layer features of ConvNets are used for face verification, then the features are compact and discriminative, but the system requires strong supervision signals from face classification training
Solution Approach 1:
The ConvNets are trained with face classification as a pre-training task, which provides strong supervision signals. The same network architecture and trained weights are then reused for feature extraction in verification, making the system multi-functional and reducing the need for separate verification-specific training.
Data Source
AI summary
Disclosed are a method and an apparatus for face verification. The apparatus comprises a feature extracting unit configured to extract HIFs (Hidden Identity Features) for different regions of faces by using differently trained ConvNets, wherein last hidden layer neuron activations of said ConvNets are considered as the HIFs. The apparatus further comprises a verification unit configured to concatenate the extracted HIFs of each of the faces to form a feature vector, and then compare two of the formed feature vectors to determine if they are from the same identity or not.